计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (19): 224-230.DOI: 10.3778/j.issn.1002-8331.1907-0214

• 工程与应用 • 上一篇    下一篇

机器人目标抓取区域实时检测方法

卢智亮,林伟,曾碧,刘瑞雪   

  1. 广东工业大学 计算机学院,广州 510006
  • 出版日期:2020-10-01 发布日期:2020-09-29

Real-Time Detection Method for Robot Target Grasping Area

LU Zhiliang, LIN Wei, ZENG Bi, LIU Ruixue   

  1. School of Computers, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2020-10-01 Published:2020-09-29

摘要:

针对目前机器人目标抓取区域检测方法无法兼顾检测准确率和实时性的问题,提出一种基于SE-RetinaGrasp神经网络模型的机器人目标抓取区域实时检测方法。该方法首先以一阶目标检测模型RetinaNet为基础提取抓取框位置及抓取角度;针对抓取检测任务采用SENet结构确定重要的特征通道;结合平衡特征金字塔设计思想,充分融合高低层的特征信息,以加强小抓取框的检测性能;在Cornell数据集上进行实验验证,结果表明该方法在取得更高检测准确率的同时,提高了抓取检测的效率,达到实时检测的要求。

关键词: 抓取区域检测, SENet结构, 平衡特征金字塔, 实时检测

Abstract:

Aiming at the problem that the current robot target grasping area detection method cannot take into account the detection accuracy and real-time performance, a real-time detection method of robot target grasping area based on SE-RetinaGrasp neural network model is proposed. Firstly, based on the one-stage target detection model RetinaNet, the position of the grasping rectangles and the grasping angle are extracted. Secondly, the SENet structure is used to determine the important feature channel in the grabbing detection task. Then, combined with the balanced feature pyramid design idea, the fusion is fully integrated. The feature information of the upper and lower layers is used to enhance the detection performance of the small grasping rectangles. Finally, the experimental verification is performed on the Cornell dataset. The results show that the method improves the detection accuracy while improving the detection accuracy, efficiency to meet real-time detection requirements.

Key words: grasping area detection, SENet structure, balanced feature pyramid, real-time detection